曹承礎臺灣大學:資訊管理學研究所黃志成Huang, Chih-ChengChih-ChengHuang2010-05-052018-06-292010-05-052018-06-292008U0001-2007200819323500http://ntur.lib.ntu.edu.tw//handle/246246/179866隨著Web 2.0與長尾效應概念的崛起，網際網路上有著越來越多的資訊可供取得，跟過去比起來，消費者現在擁有更多選擇。然而，擁有選擇眾多產品的權利卻未必是件好事。好在，設計用來協助使用者找尋適合自己產品的推薦系統提出了一個新的解決方式。許許多多的學者也投入推薦系統相關的研究，其中協同過濾、內容過濾、知識與個人特徵等四種推薦方法儼然成為這個領域的主流，這些方法各自有各自的優缺。所以，將這些方法融合成單一推薦系統的混合法便因此被衍生出來。在這個研究當中，我們將協同過濾法與個人特徵法以特徵結合的方式混合，並借助資料探勘中計算相似度的演算法來決定使用者之間的相似度。同時，也將內容過濾與知識法用特徵增強的方式整合，來找出類似的產品。由於化妝保養品本身缺乏統一標準、高單價與高涉入等特性，我們用它來作為推薦概念的驗證。根據實驗的結果顯示，我們的作法跟傳統Pearson相關係數的使用者協同過濾法有著相同的預測準確度，卻有更高的分類與排序的準確度。實驗的參與者針對系統的可用性、推薦新產品的能力、採用率與整體滿意度皆有著令人滿意的結果。基於上述幾點理由，我們認為此項研究最大的貢獻在於提出一個推薦產品或服務更新更好的作法。With the recent rise of Web 2.0 concepts and the advent of a long tail economy, more and more content can be obtained though the Web. Consumers now have much more alternatives than ever before. Nonetheless, the plenty of choices is itself a blessing and a curse. las, the recent birth of the recommender system, which aims to find the items that a specific user might be interested in, provides us with a new remedy. So much effort has been devoted to this area of research and four different approaches; namely, collective filtering, content filtering, knowledge-based, and demographic; have become the four major recommendation techniques. Each has its own pros and cons. As a result, one of the branches of recommender system research is to blend these mechanisms into a single hybrid. n this paper, we extrapolate the feasibility of the feature combination hybrid method by merging the collective filtering and demographic techniques. Meanwhile, an idea from data mining field was borrowed to develop a new way in computing the similarity between users. We also combine the content filtering and knowledge-based by using the feature augmentation hybrid method to filter out similar products. Skin care products are chosen to be our proof-of-concepts due to their often semi-standard product nature, their general high price, and the high user involvement in the purchasing process. he empirical result demonstrates that our approach has similar prediction accuracy as the Pearson correlation metric, proven to be the most accurate one in terms of mean absolute error, while at the same time having higher classification and ranking accuracy. The participants also reveal having satisfactory level of system usefulness, novelty, adoption and satisfaction. It is therefore our strong believe that our contribution lies in the building of a novel and improved approach for recommending goods and services.1. Introduction 1.1. Background 1.2. Motivations 2.3. Objectives 4.4. Organization 4.5. Contributions 5. Literature Review 6.1. Recommender Systems 6.2. Traditional Similarity Measures 16.3. Discussion 18. System Design & Implementation 23.1. Scenario 23.2. System Overview 24.3. System Components 26. Experiment & Evaluation 27.1. Evaluation Criteria 27.2. Baseline Algorithms 29.3. Experiment Procedure 29.4. Hypothesis Testing & Analysis 35.5. Conclusion 40. Limitation 42. Future Work 44eference 46application/pdf1754710 bytesapplication/pdfen-US混合式推薦系統化妝品推薦特徵增強特徵結合hybrid recommender systemcosmetics recommendationfeature augmentationfeature combinationhttp://ntur.lib.ntu.edu.tw/bitstream/246246/179866/1/ntu-97-R95725041-1.pdf